Sun Jingxiang ; Wang Xuan ; Shi Yichun ; Wang Lizhen ; Wang Jue ; Liu Yebin

Abstract

Existing 3D-aware facial generation methods face a dilemma in quality versuseditability: they either generate editable results in low resolution orhigh-quality ones with no editing flexibility. In this work, we propose a newapproach that brings the best of both worlds together. Our system consists ofthree major components: (1) a 3D-semantics-aware generative model that producesview-consistent, disentangled face images and semantic masks; (2) a hybrid GANinversion approach that initialize the latent codes from the semantic andtexture encoder, and further optimized them for faithful reconstruction; and(3) a canonical editor that enables efficient manipulation of semantic masks incanonical view and product high-quality editing results. Our approach iscompetent for many applications, e.g. free-view face drawing, editing, andstyle control. Both quantitative and qualitative results show that our methodreaches the state-of-the-art in terms of photorealism, faithfulness, andefficiency.